huggingface.co web signal

Hy-Nomas Ships 8B Interactive World Model on a Single RTX 4090

5 sources tracking this story

TL;DR

  • A single RTX 4090 running INT8 quantization and CPU offloading covers the model, expanding access well beyond the 8-GPU reference configuration.
  • A persistent KV cache carries scene context across turns, enabling continuous multi-segment world generation without re-rendering prior frames.
  • Tencent published no technical report at launch; all performance figures come from proprietary benchmarks, blocking independent replication.

Hy-Nomas, an 8B diffusion video world model, generates a 100-second clip at 1280×704 and 24 FPS on a single RTX 4090, with peak GPU memory of 20.53 GiB.

The release, posted on Hugging Face, pairs that inference recipe with pairwise human-evaluation numbers against three baselines. On the multi-turn free-form text benchmark, Hy-Nomas beats LingBot-World v2 75.0% of the time on prompt following and 68.3% on physical plausibility. On the interleaved text-and-camera benchmark, the numbers against the adapted LingBot-World v2 are 75.0%, 73.3%, and 26.7% across prompt following, physical plausibility, and camera control respectively.

Against Google DeepMind's Genie 3 the picture flips. On spatial memory, which tests object permanence after a viewpoint revisit, Hy-Nomas wins 13.3% of comparisons, ties 56.7%, and loses 30.0%. The paper names this as an area for improvement.

The abstract's core design claim: "Our central design aligns structured video captions with action-aware causal attention: descriptive context remains accessible across segments, while action instructions are restricted to their intended temporal windows." The 8B backbone is a 54-block MMDiT paired with a Qwen2.5-VL-7B text encoder, and Hy-Nomas-Bench bundles 280 test cases across four benchmarks, hosted at huggingface.co/datasets/tencent/Hy-Nomas-Bench.

The 4090 recipe is not trivial: INT8 quantization, sliding-window attention, and CPU offloading sit behind the 20.53 GiB figure. Weights, training code, and a LoRA adaptation workflow ship together, and the paper demonstrates cinematic and robotics LoRAs trained on roughly 100 and 128 clips respectively. Google DeepMind debuted Genie 4 the day before.

What others are reporting

Coverage cluster as of 24h after publish

  1. GIGAZINE Read →

    Consumer hardware angle: confirms the 8B model runs on a single RTX 4090 via INT8 quantization and CPU offloading, with direct links to weights, code, and browser demo.

    The model is available as an open-source model and can be run locally on consumer hardware using int8 quantization and CPU offloading.
  2. VP Land Read →

    Flags that Tencent had not published a technical report at release and relies solely on proprietary benchmarks, directly questioning the gap between announcement and independent validation.

    you can keep adjusting subjects, actions, style, or camera movement across turns
  3. ArtRealmAI Read →

    Developer-focused breakdown covering actual JSON inference structure, honest RAM and latency trade-offs, and explicit clarification that the model is not a physics engine and has no queryable state.

    Hy-Nomas takes sentences. Start a scene, then on the next turn say 'glowing flowers open beside the path.'
  4. AI Polix Read →

    Technical breakdown of the dual-stream MMDiT architecture and Segment Forcing training method, plus early adoption signals: 186 GitHub stars and 60K views by October 26.

    Each segment is produced by a diffusion transformer conditioned on the latest instruction, an optional camera trajectory and a KV cache carrying prior context.